Skip to the content.

The notes the builds were planned from. They are working documents: dated, written in the middle of the work, and sometimes superseded by a later measurement. Where a note and a build page disagree, the build page is later and governs. Start with the three files marked start here.

Every file below is PLI Labs’ own note, copied from our private research archive. Each says at its top where it came from and what was changed for publication: machine names, private file paths, network addresses, session names and the operator’s name were replaced with generic terms; links to unpublished files became plain text; a few lines that named the author of a note or quoted an instruction were shortened. No figure and no finding was changed. Third-party work is summarised and linked in these notes, never copied: follow the links to the originals.

Volta architecture and the kernel rules

File What it holds Date
kernels/KERNEL-RESEARCH-EXPLAINER.md Start here. What a GPU kernel is, why decode is limited by memory and hand-offs, how to build one well, a glossary. Plain terms 2026-11-02
kernels/KERNEL-RESEARCH-CLAUDE-NOTES.md Start here. Twelve binding rules for anyone (or any coding agent) writing a kernel for sm_70 or the DGX Spark, each with its source 2026-11-02
kernels/KERNEL-RESEARCH-REPORT.md Start here. The full survey: the roofline, memory hierarchy, landmark kernels, hardware generations from Volta to Blackwell, AI-written kernels, and its reference list 2026-11-02
kernels/CUSTOM-KERNEL-OPPORTUNITIES-20261109.md The ranked list of kernel opportunities written for the eight-card NVFP4 build 2026-11-09
kernels/KOPT1105-ROOFLINE.md A CPU-only roofline and census analysis of the prefill expert kernel: where padded work comes from 2026-11-05
kernels/KOPT1105-UPSTREAM.md What upstream offered for the same kernel 2026-11-05
kernels/UPSWEEP-ATTN-PRIOR-ART.md Prior art for attention kernels on sm_70: what exists upstream and what it assumes 2026-11-01
volta/v100-skinny/FOUNDATION.md Our review of v100-skinny (Dennis Vertzyas’ kernels) as a foundation: what it proves and what it leaves open 2026-09-16
volta/wiki/05-new-archs-msa-vs-csa.md Two newer attention designs in plain terms, and why one is easy on Volta 2026-07-07

Where the time goes (measured)

File What it holds Date
kernels/PROF1103-REPORT.md The first profile of a decode round on eight cards: 32 ms split by stage and kind of work, and why one upstream pull request did not transfer 2026-11-03
kernels/PROF1103b-REPORT.md The second profile, after the first levers 2026-11-03
kernels/PROF1-report.md The earlier profile of the EXL3 engine’s round 2026-11-01
kernels/TOK1103-REPORT.md Tokens-per-second levers found in v100-skinny and 1Cat-vLLM, with upstream numbers and our estimates 2026-11-03
LANE-RESEARCH.md in Volta Max uPLIft Where the time of a decode round goes on the eight-card NVFP4 build. Published in that build’s repository; not copied here 2026-11-09
eight-card/topk-full-20261110.md The full cost of the index top-k kernels per decode round, measured on one card of a four-card test machine 2026-11-10

The twelve-card time budget is in that build’s repository and summarised in ../optimizations/volta-12x-amplify.md.

Upstream sweeps

File What it holds Date
kernels/UPSCAN1105-REPORT.md A read-only sweep of public work for ideas that port to this build, ranked 2026-11-05
kernels/UPSCAN1105-ONECAT.md The same for 1Cat-vLLM alone 2026-11-05
kernels/UPSCAN1105-SUPPLEMENT.md Its supplement 2026-11-05
volta/findings/V-C-speedups-untried.md Speed-ups not yet tried on V100, June 2026: an early survey, written before the DeepSeek work. Read it for the options, not for the numbers 2026-06
volta/findings/V-D-community-china.md What the wider community had measured on V100s, including Chinese-language sources, June 2026 2026-06

The first DeepSeek-V4.1 port (October 2026, behind 1Cat-vLLM #1031)

The record of the first twelve-card port: three stages of four cards, the official checkpoint, no CUDA graphs. It ended at about 3 tok/s and an overall FAIL of its exactness gates, and it is direct prior art for Volta 12X AmPLIfy.

File What it holds
volta/deepseek-v4.1-flash/PORT_DESIGN.md The design of the port: package layout, contracts between parts, the stage cuts, the test plan
volta/deepseek-v4.1-flash/MODEL_ARCHAEOLOGY.md The model read from its reference code: what every part does
volta/deepseek-v4.1-flash/V4_TO_V41_MODEL_DIFF.md What changed from DeepSeek-V4 to V4.1
volta/deepseek-v4.1-flash/1CAT_ARCHAEOLOGY.md 1Cat-vLLM read as a codebase: what it had for DeepSeek-V4 and what V4.1 needed
volta/deepseek-v4.1-flash/TENSOR_CENSUS.md Every tensor of the checkpoint by kind, format and size
volta/deepseek-v4.1-flash/ENGRAM.md The Engram n-gram memory: tables, hashing, how it is served
volta/deepseek-v4.1-flash/CHECKPOINT_CANDIDATES.md Which published checkpoints could be served on Volta, and why
volta/deepseek-v4.1-flash/RANGE_AUDIT.md The numeric ranges of activations, for FP16 safety
volta/deepseek-v4.1-flash/NLL_GAP.md How much of a quality gap came from the port’s precision choices (answer: almost none)
volta/deepseek-v4.1-flash/E2E_FINAL.md The final end-to-end campaign and its FAIL. Three raw API responses were removed: generated model text is not published here

Not copied here, and where to find it

Material Why not Where
The v100-skinny kernels, benchmarks and results Dennis Vertzyas’ work https://github.com/dnv2003/v100-skinny
The 1Cat-vLLM SM70 kernels and pull-request texts their authors’ work https://github.com/1CatAI/1Cat-vLLM; every pull request we used is linked in ../CREDITS.md
Our production kernels (v41k, the index scorers, the prefill expert GEMM) already public as patches Volta sPLIt, engine/production
The RB 4 prompt-reading kernel already public as source Volta Max uPLIft, engine/kernels/prefill-rb4
The plans of the eight-card NVFP4 build already public Volta Max uPLIft, docs/plans
The lever log of the eight-card build already public Volta sPLIt, docs/levers.md
The June 2026 notes on the serving stack and on serving larger models they describe one machine’s inventory in detail; their conclusions were published at the time pli-labs, research/v100
The long working logs of individual kernel experiments, the kernel campaign on the test machine, the hardware inventory working logs tied to one machine. Their results are in the build pages ../optimizations/
The twelve-card build’s measurements, history, kernel plan and time budget published in the build’s own repository Volta 12X AmPLIfy; summarised in ../optimizations/volta-12x-amplify.md